Generalized linear mixed models with Gaussian mixture random effects: inference and application (Q2293384)

This paper investigates generalized linear mixed models (GLMM) with Gaussian mixture random effects for clustered data. A penalized EM algorithm is proposed. Sequential locally restricted likelihood ratio tests are set to decide the number of components in the mixture model. The proposed inference procedure can be used to test if there are any clusters among the units and thus examine the goodness-of-fit of the classic GLMM assuming homogeneous Gaussian random effects. When the proposed tests are used sequentially, the procedure can automatically determine the number of mixture components. The method is applied on a kidney transplant centers evaluation, using the U.S. Organ Procurement and Transplantation Network database. The patient 5-year post-transplant survival status is modeled using a GLMM where the random effect for a transplant center follows a finite Gaussian mixture distribution. The empirical Bayes approach is then applied to classify the transplant centers using the fitted Gaussian mixture model, while controlling the false discovery rate when screening for transplant centers with non-standard performance. Simulation studies are also performed.